A research project (in collaboration with Meta AI) on improving the planning and long-horizon reasoning of large language models.

I trained a text encoder together with a Graph Neural Network (GNN) in PyTorch to retrieve graph-structured plans, which the LLM then uses to reason over multi-hop question-answering tasks. Grounding generation in an explicit, retrieved plan improves complex reasoning compared to free-form prompting.